MAGMA: Machine learning Automatic picker for Geothermal Microseismicity
Analysis for practical procedure to reveal fine reservoir structures
Abstract
In geothermal development, microseismic monitoring is important
technique to monitor the various phenomena in the reservoirs throughout
location, activity, and magnitude of microseismicity. Picking P- and S-
wave arrivals accurately from seismic data is inevitable process for
subsequent seismic analyses. However, the manual phase picking is a
time- and cost-consuming process and several automatic pickers still
requires considerable quality checks and corrections by human analysts.
Automatic pickers based on deep learning have recently been developed
for natural earthquake analysis, whose accuracy has been confirmed to be
comparable to that of human analysts. These phase pickers were mainly
trained using natural earthquakes recorded by regional seismic networks.
However, seismic networks and events in geothermal fields have features
that differ from those of natural earthquakes. In such fields, seismic
events with very low magnitudes occur immediately under the seismic
network and are sometimes triggered by fluid activity. Therefore, the
direct application of the existing deep learning phase pickers to such
seismic networks may have difficulty. Here, we focus on developing a
deep learning model specialized for local seismic networks in geothermal
fields. We used microseismic data from four representative enhanced
geothermal and hydrothermal fields and trained the model with deep
learning. Based on the developed model, the hypocenter distribution was
determined using continuous seismic waves in the Okuaizu geothermal
field, Japan. These procedures were performed automatically without
manual operations and we propose them as MAGMA: Machine learning
Automatic picker for Geothermal Microseismicity Analysis. Subsurface
fine structures were then revealed by relocating the hypocenters using a
double-difference algorithm. The same procedures for the same data were
then conducted using a deep learning model that trained by other field
data, and the equivalent structures were successfully reveled. Thus,
MAGMA is applicable to new fields even when data is lacking, such as
green fields.